EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

fix: second_per_grid_ts should be used to get mrope position (#3682)

Browse Source
This commit is contained in:
Mick
2025-03-18 09:12:38 +08:00
committed by GitHub
parent 98be3bd306
commit d373a48c98
8 changed files with 93 additions and 69 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
python/sglang/srt/model_executor/model_runner.py
View File

@@ -258,10 +258,12 @@ class ModelRunner:
if self.model_config.hf_config.architectures == [
"Qwen2VLForConditionalGeneration"
] or self.model_config.hf_config.architectures == [
"Qwen2_5_VLForConditionalGeneration"
]:
# TODO: qwen2-vl does not support radix cache now, set disable_radix_cache=True automatically
# TODO: qwen2-vl series does not support radix cache now, set disable_radix_cache=True automatically
logger.info(
"Automatically turn off --chunked-prefill-size and disable radix cache for qwen2-vl."
"Automatically turn off --chunked-prefill-size and disable radix cache for qwen-vl series."
)
server_args.chunked_prefill_size = -1
server_args.disable_radix_cache = True